US blocks Anthropic's best AI over fears of China stealing hacking secrets

The panic started when Anthropic suddenly turned off its most powerful models, Mythos and Fable 5.

Officially, the drama began because Amazon engineers found a massive security loophole. When they asked Fable 5—which was supposed to be a highly secure version of Mythos—to check some code for bugs, the AI didn't just find errors; it pointed out exactly how to exploit those vulnerabilities to hack the system.

This oopsie triggered the White House. Washington went into full red-alert mode, fearing that a China-backed hacker group might have already poked around the system. The American officials worry that Beijing could use these models to map out exploits in US critical infrastructure, or worse, steal the AI's intelligence via distillation—a clever trick where a cheaper Chinese model learns directly from the answers of the American super-AI, skipping years of expensive R&D.

Anthropic is trying to cool down the drama, claiming the White House never mentioned Chinese hackers in their export control chats, and besides, their services are already geoblocked in China. Even David Sacks, a prominent tech advisor, only ranted on X about the jailbreak itself without bringing up geopolitical espionage. The company insists this vulnerability is incredibly narrow and that competitors like GPT-5.5 can be tricked into doing the exact same thing.

It seems the ultimate defense mechanism of the free world is just pulling the power plug whenever the math gets too smart. Watching superpowers fight over digital ghosts while the actual tech gets locked in a basement is peak 21st-century comedy.

Source: Semafor